Random

• 🇨🇳 My Chinese name is 刘佳伟.
• 🇩🇪 Ich kan Deutsch sprechen aber nur ein bischen.
• I enjoy cooking, building PCs, CS2, anime, hiking, and boba tea 🧋.
• 🌽 UIUC Beamer-style Google slides.

Quotes

“Simplicity is the prerequisite for reliability.” — Edsger W. Dijkstra

“Given enough eyeballs, all bugs are shallow.” — Linus’ law

Compute and random walking

At OpenAI, I prompt Codex to train agents and my work has been integrated to reasoning models from GPT-5.2 to 5.4.

Before joining OpenAI, I spent four wonderful years at UIUC to pursue my PhD, advised by Lingming. I was mostly thinking about ideas in the space between machine learning and programming systems. A core theme of my work explores how to automate software engineering by automating testing, in sense of both finding bugs in complicated software and evaluating LLMs. During my PhD, I trained a bunch of coding models, built popular evaluators, and created SOTA programmatic fuzzers for AI systems. I also interned at AWS, Google, and OctoAI, where I studied code preference, improved TPU reliability, and built a pattern language for AI compilers. I also led the UIUC team to win the championship of Amazon Nova AI Challenge 2025.

Before that, I random-walked a ton of things during my undergrad at Tongji University. In my 1st year, I was into mathematical modeling that I made reference answers for our new textbook and won a national prize. I transfered to CS in my 2nd year, where I used to be a “language laywer” of modern C++ after getting mid-trained over cppreference. I co-led the robotic vision efforts in our prize-winning university team (SuperPower) in RoboMaster Robotics Competition, hosted by DJI. I finally seriously started my research journey with Jinyang on programmable video analytics and Luo on efficient 2D pose estimation. I also worked on model serving before it was cool at ByteDance AI Labs and GNNs at Alibaba DAMO Academy.

Before undergrad, I knew nothing about coding and only used computers to play video games.